Ex OpenAI CTO is fixing one of AI biggest problems

@sarainwondertech

12.5K followers · usually 1.4K views a post

Views

3.6K

Likes

Comments

SoundTikTok sound · original sound - sarainwondertech

Promoted product

Thinking Machines Lab

Connectionism: Research Blog by Thinking Machines Lab

3 videos2 creators

Why it worked

The video effectively explains a complex AI concept using clear analogies and visuals, addressing a common concern about AI replacing humans by offering a collaborative alternative. The creator's energetic delivery and the focus on a prominent figure (Mira Murati) likely contributed to its engagement.

Summary

The video explains that current AI models like ChatGPT are designed with the assumption of full delegation, aiming to replace humans. However, the former CTO of OpenAI, Mira Murati, and her startup Thinking Machines Lab are developing an "interaction model" that focuses on real-time collaboration between humans and AI. This new approach splits the AI into a fast, always-present interaction model and a powerful background model for heavy lifting, aiming to create a more conversational and collaborative AI experience.

Structure

1Problem with current AI: designed for delegation, not collaboration.
2Mira Murati's solution: an "interaction model" for human-AI collaboration.
3

Original caption

Mira Murati's startup Thinking Machines Lab just dropped the technical details for an "interaction model", and it's a direct shot at the autonomy consensus the rest of the industry is racing toward. The thesis: AI labs are designing for a future where you're not in the loop. Every architectural choice in ChatGPT, Claude, Gemini assumes you delegated and walked away. Thinking Machine Lab is going the opposite direction: building AI trained natively to *collaborate* with you in real time, not replace you. The architecture is genuinely different. Every 200 milliseconds, the model processes whatever's happening — your voice, your typing, what's on your screen — and produces an output (speech, text, an action, or silence). No more stitching together speech-to-text + LLM + text-to-speech + tool calls. One model, trained on the interaction itself. And to keep both real-time speed AND deep intelligence, they split the system into two models: a small fast one that stays with you, a bigger slow one that handles heavy reasoning in the background. Research preview only, coming in the next few months. Whether the architecture holds up across longer real-world use is still an open question. But the philosophical break is real, and it's the first one any major lab has made. #airesearch #thinkingmachineslab

More from @sarainwondertech

On-screen text

Ex OpenAI CTO is fixing one of AI biggest problems design v1.0 human AI agent THIS IS THE BIGGEST PROBLEM IN A I TODAY AND IT'S INTENTIONAL CHATGPT GEMINI CLAUDE BUILT UNDER THE ASSUMPTION THAT THE AI AGENT SHOULD BE DOING EVERYTHING AND YOU SHOULDN'T BE THERE AT ALL THEY ARE BASICALLY BUILT WITH THE GOAL TO REPLACE YOU EXCEPT THIS IS A FANTASY AI AGENTS NEED SUPERVISION AND CANNOT BE LEFT ALONE THEY WILL ACTUALLY WORK 10 X BETTER DESIGNED TO COLLABORATE WITH AND EVEN IF KNOWN PROBLEM NO ONE IS FIXING THIS EXCEPT THIS WOMAN FORMER CTO OF OPEN AI MIRA MURATI HER STARTUP THINKING MACHINES LAB CALLS THIS AN INTERACTION MODEL BECAUSE IT'S DESIGNED TO COLLABORATE WITH YOU IN REAL TIME LET ME SHOW SO WITH A NORMAL LLM WHAT TALKING TO CHAT ACTUALLY LOOKS LIKE FEELS LIKE HAVING A CONVERSATION IT'S ACTUALLY MORE LIKE A WALKIE TALKIE OF DIFFERENT TOOLS GLUED TOGETHER AND IT'S SEQUENTIAL BECAUSE THE ASSUMPTION IS THAT YOU'RE DELEGATING TO AI NOT WORKING WITH BUT WHAT MIRA MURATI IS BUILDING IS LESS WALKIE TALKIE AND MORE LIKE A PHONE CALL SEQUENCE LIKE THIS ONE WHILE YOU TALK, THE MODEL TICKS - 5 TIMES A SECOND. 00:28:47 speech silence. action. THE MODEL PROCESSES EVERYTHING THAT IS GOING ON EVERY 200 MILLISECONDS EVERY 200ms: A FULL CYCLE. voice typing screen 200ms 200ms 200ms 200ms speech / text / action / silence input, output after tick. SO EVERY 200ms THE MODEL TAKES IN WHATEVER IT'S HAPPENING WHAT YOU'RE SAYING WHAT YOU'RE TYPING SCREEN AND RETURNS AND THE OUTPUT CAN BE TALKING TEXTING TAKING AN OR SILENCE DOING NOTHING AND IT REPEATS LIKE THIS EVERY 200 MILLISECONDS SO THAT IT A STREAM THAN A SEQUENCE THE END RESULT FEEL LIKE YOU'RE HAVING AN ACTUAL THIS IS POSSIBLE BECAUSE THE MODEL IS NOT RELYING ON EXTERNAL TOOLS BUT IT'S TRAINED ON THE INTERACTION ITSELF CATCH A MODEL SO FAST IS USUALLY NOT THAT INTELLIGENT THE SPLIT. interaction model you small, fast, always with you. hand-off background model reasoning code research tools big, slow, heavy thinking. fast model thinks. they coordinate. SO THE SOLUTION IS SPLITTING THE MODEL IN TWO SMALLER AND FASTER THAT ALWAYS STAYS AND A SECOND BIGGER AND WORKS IN THE BACKGROUND DOING ALL THE HEAVY LIFTING THIS MODEL WILL BE AVAILABLE IN THE COMING MONTHS THAT WE WILL MORE OF THIS HI MY NAME I SPENT THE PAST 10 YEARS BUILDING AI PRODUCTS AND IF YOU SENSE OF THE NEWS AND RESEARCH HIT FOLLOW

Transcript

This is the biggest problem in AI today and it's intentional. ChatGPT, Gemini, Claude, they are all built under the assumption that the AI agent should be doing everything and you shouldn't be there at all. They are basically built with the goal to replace you. Except this is a fantasy. AI agents need supervision and cannot be left alone. They will actually work 10x better designed to collaborate with you. And even if it's a known problem, no one is fixing this. Except this woman. Former CTO of OpenAI, Mira Murati. Her startup, Thinking Machines Lab, calls this an interaction model because it's designed to collaborate with you in real time. Let me show you the difference. So with a normal LLM, what talking to ChatGPT actually looks like. It feels like having a conversation. It's actually more like a walkie talkie of different tools glued together and it's sequential. Because the assumption is that you're delegating to AI, not working with it. But what Mira Murati is building is less walkie talkie and more like a phone call. Instead of a sequence like this one, while you talk, the model ticks - 5 times a second. The model processes everything that is going on every 200 milliseconds. Every 200ms: A full cycle. So every 200ms, the model takes in whatever it's happening, what you're saying, what you're typing, what's on your screen and returns and the output can be talking to you, texting, taking an action or silence doing nothing. And it repeats like this every 200 milliseconds. So that it feels more like a stream than a sequence. The end result is that you feel like you're having an actual conversation. This is possible because the model is not relying on external tools, but it's trained on the interaction itself. But there's a catch. A model so fast is usually not that intelligent. So the solution is splitting the model in two. One smaller and faster that always stays with you. And a second bigger and works in the background, doing all the heavy lifting. This model will be available in the coming months. And we will see more of this. Hi, my name is Sara. I spent the past 10 years building AI products. And if you want to make sense of the news and research, hit follow.